Files
OmniRoute/open-sse/executors/gemini-web/capabilities.ts
2026-08-08 12:47:48 -03:00

122 lines
5.5 KiB
TypeScript

/**
* Request-contract guards for the Gemini Web executor (#9356).
*
* gemini-web is not an API client. It launches Playwright, types ONE flat
* prompt string into the gemini.google.com `.ql-editor` contenteditable,
* presses Enter, and captures the first `StreamGenerate` response off the page
* (see ../gemini-web.ts). There is no JSON request body on the wire, which
* makes two OpenAI controls structurally impossible to honor:
*
* • `reasoning_effort` — no field exists to carry a thinking budget. Unlike
* deepseek-web or perplexity-web, which post a real payload and can flip a
* `thinking_enabled` flag or swap the model preference, there is nothing
* here to set.
* • forced `tool_choice` — the tools support gemini-web does have is the
* prompt-emulation shim (`translator/webTools.ts`, #7286): it ASKS the
* model, in prose, to answer with `<tool>{...}</tool>` and parses whatever
* comes back. That is best-effort by construction. "required" / "any" /
* a named function is a GUARANTEE, and a prompt cannot make one.
*
* Before this module both were accepted and quietly ignored, so an agent got a
* 200 with `finish_reason: "stop"`, no `reasoning_content`, and `tool_calls: []`
* and concluded its requirements had been met (#9356). Failing the request is
* the honest answer: the caller can drop the control, or route to a model that
* actually implements it.
*
* Deliberately NOT rejected — these are already satisfied or already work:
* • `reasoning_effort: "none" | "minimal"` — asking for as little reasoning as
* possible is something a non-thinking provider trivially complies with.
* • `tool_choice: "auto" | "none"` and plain `tools[]` — the #7286 emulation
* path, which several shipped combos depend on (#5240, #8488). Untouched.
*
* Pure and dependency-free so the whole contract is unit-testable without a
* browser.
*/
/** `error.code` on every compatibility rejection raised here. */
export const GEMINI_WEB_UNSUPPORTED_CONTROL_CODE = "unsupported_control_for_provider";
/** Effort levels a non-thinking provider already complies with. */
const SATISFIED_EFFORT_LEVELS = new Set(["none", "minimal"]);
/** `tool_choice` strings that demand a tool call rather than merely offering one. */
const FORCING_TOOL_CHOICE_STRINGS = new Set(["required", "any"]);
/** `tool_choice: { type }` values that pin the model to a specific/any tool. */
const FORCING_TOOL_CHOICE_TYPES = new Set(["function", "tool", "any"]);
export interface GeminiWebCapabilityViolation {
/** Which request field could not be honored. */
param: "reasoning_effort" | "tool_choice";
/** Client-facing explanation — already safe to put in a response body. */
message: string;
}
function normalizeString(value: unknown): string | null {
return typeof value === "string" && value.trim().length > 0 ? value.trim().toLowerCase() : null;
}
/**
* True when `tool_choice` demands a tool call. Covers the OpenAI strings
* ("required"), the Anthropic-flavored ones the translators also emit ("any"),
* and the object forms that name a function or force any tool. "auto" / "none"
* and every unrecognized shape are treated as non-forcing — this guard only
* blocks contracts it is certain gemini-web cannot keep.
*/
export function isForcingToolChoice(toolChoice: unknown): boolean {
const asString = normalizeString(toolChoice);
if (asString) return FORCING_TOOL_CHOICE_STRINGS.has(asString);
if (toolChoice && typeof toolChoice === "object" && !Array.isArray(toolChoice)) {
const type = normalizeString((toolChoice as Record<string, unknown>).type);
return type !== null && FORCING_TOOL_CHOICE_TYPES.has(type);
}
return false;
}
/** True when `reasoning_effort` asks for MORE thinking than "none at all". */
export function requestsThinkingBudget(reasoningEffort: unknown): boolean {
const effort = normalizeString(reasoningEffort);
if (effort === null) return false;
return !SATISFIED_EFFORT_LEVELS.has(effort);
}
/**
* Inspect an OpenAI-shaped request body for controls gemini-web cannot honor.
* Returns the first violation found, or `null` when the request is servable.
*
* `reasoning_effort` is checked before `tool_choice` only for determinism; a
* request carrying both is rejected either way.
*/
export function checkGeminiWebUnsupportedControls(
body: Record<string, unknown> | null | undefined
): GeminiWebCapabilityViolation | null {
if (!body || typeof body !== "object") return null;
if (requestsThinkingBudget(body.reasoning_effort)) {
return {
param: "reasoning_effort",
message:
'Model provider "gemini-web" does not support "reasoning_effort". It drives the ' +
"gemini.google.com web UI through a typed prompt and has no thinking-budget control " +
'to set, so any effort above "minimal" would be silently ignored. Remove ' +
'"reasoning_effort" (or send "none"/"minimal") or route to a reasoning-capable model.',
};
}
if (isForcingToolChoice(body.tool_choice)) {
return {
param: "tool_choice",
message:
'Model provider "gemini-web" cannot guarantee a forced tool call. Its tool support is ' +
"prompt-emulated — the model is asked to emit a tool block and may answer with prose " +
'instead — so "tool_choice" values that require one ("required", "any", or a named ' +
'function) cannot be honored. Use "auto" to keep best-effort tool calling, or route to ' +
"a model with native function calling.",
};
}
return null;
}